On-Device Web AI100% Private (No API Keys)

Browser MCP Studio: Run On-Device AI Tools

Run, test, and debug Model Context Protocol servers directly inside Google Chrome with on-device Gemini Nano, WebGPU, and frontier models. Zero local CLI installs, zero server latency, and real-time JSON-RPC 2.0 protocol inspection.

Active MCP Server:
Tools (3):list_issuescreate_issuesearch_repositories
Resources (2):
Prompts (1):
mcp-agent-cli:~/github
⚡ MCP CODEX CLI v2.6|stdio/sse wire operational
● Online

Connected to GitHub with 3 tools registered. Run commands below or execute a 1-click autonomous workflow:

▶ 1-Click Autonomous Workflows (ReAct Loop):
AGENTlocal-execution-thread
05:00 PM
⚡ **MCP CODEX AGENT TERMINAL v2.6 Ready** Connected to **GitHub MCP Server** via JSON-RPC 2.0 wire. Type any command below or click a quick action to dispatch tools.
Quick:
mcp:github$

MCP Protocol Inspector

Total Roundtrip0 ms
Tool Latency0 ms
Local Inference0 ms

No active tools/call yet.

Send a message to watch live JSON-RPC frames.

Interested in GitHub config?View Full Server Docs
Technical Architecture

How In-Browser Model Context Protocol Works

Traditionally, connecting an LLM to Model Context Protocol (MCP) required running local background processes (via Node.js, Python, or Docker) and piping stdio streams through desktop host applications like Claude Desktop or Cursor. Browser AI Studio brings the full MCP agent loop directly into the web browser using Chrome’s native Prompt API, WebGPU, and client-side HTTP/SSE transports.

1. Native Gemini Nano (Chrome)

Executes via Chrome's experimental window.ai.languageModel Prompt API. Runs a 3.2B parameter quantized Gemini Nano model directly on local CPU or NPU with 0 API tokens billed and zero network transit.

2. In-Browser WebGPU Inference

Powers client-side Small Language Models (SLMs) compiled to WebAssembly (Wasm) and executed via WebGPU compute shaders inside a dedicated Web Worker, preventing UI frame drops during agent planning loops.

3. Direct Frontier BYOK

Connect your private API keys for Google Gemini 3.8 Flash, Anthropic Claude 3.7 Sonnet, OpenAI GPT-5, or local Ollama. Keys remain in private browser localStorage and make direct outbound fetch calls without an intermediary proxy.

Feature Matrix

In-Browser MCP Studio vs Desktop Clients

See how testing MCP tools directly inside Google Chrome compares against traditional desktop AI clients and command-line interfaces:

Feature / CapabilityBrowser AI StudioClaude DesktopCursor IDETerminal CLI
Installation RequirementZero (Runs in tab)Native App InstallerFull IDE InstallNode / Python / uv
Native On-Device Model (Gemini Nano)✓ Supported (Prompt API)✗ Cloud only✗ Cloud only✗ Cloud only
Real-Time JSON-RPC 2.0 Inspector✓ Live Wire Inspector✗ Hidden background logs✗ HiddenManual stdout capture
Remote SSE / HTTP Transports✓ Direct CORS / SSERequires wrapper bridgeRequires bridge✓ Native support
Local Privacy & Key Security100% Client-SideLocal host processCloud proxy routeLocal host process
Subscription Cost$0 Free Tier / Local$20 / month (Pro)$20 / month (Pro)Pay-per-token API
Verified Services

Explore Pre-Configured MCP Services

Test and inspect standard tool schemas from popular Model Context Protocol integrations directly in this studio:

GitHub API

Docs

Query issues, review pull requests, and search open-source repositories via GitHub REST/GraphQL.

Tools: list_issues, create_issue, search_repositories

PostgreSQL Database

Docs

Inspect schema definitions, explain query execution plans, and run read-only SQL queries.

Tools: query, list_tables, describe_table

SQLite Embedded

Docs

Lightweight embedded database inspection with zero remote infrastructure required.

Tools: read_query, describe_table

Brave Web Search

Docs

Real-time web browsing and information retrieval without tracking or search advertising bias.

Tools: brave_web_search

Local Filesystem

Docs

Safely read source files, explore project trees, and inspect sandboxed application artifacts.

Tools: read_file, list_directory

Slack Workspace

Docs

Orchestrate team communication, alert notifications, and multi-channel agent collaboration.

Tools: list_channels, post_message
Frequently Asked Questions

Frequently Asked Questions About Browser MCP

Can I run Model Context Protocol (MCP) servers directly in a web browser?▼

Yes. Browser AI Studio enables running Model Context Protocol tools client-side using Chrome Built-in AI (Gemini Nano via the Prompt API), WebGPU in-browser models, or direct client-side frontier AI (BYOK for Gemini 3.8, Claude 3.7 Sonnet, GPT-5, and Ollama). Zero Node.js or Python daemons required.

How do I enable Chrome Gemini Nano Prompt API for MCP?▼

Open Google Chrome (version 128 or higher) and navigate to chrome://flags. Enable 'Prompt API for Gemini Nano' and 'Enables optimization guide on device'. Restart Chrome, visit chrome://components, and click 'Check for update' on Optimization Guide On Device Model until fully downloaded.

Does Browser AI Studio require installing local CLI tools or Docker?▼

No. The entire agent loop and Model Context Protocol client execute directly in your browser tab. For public servers and mock schemas, everything runs client-side. For custom servers, you can connect directly to any remote HTTP or Server-Sent Events (SSE) MCP endpoint with CORS enabled.

Are my API keys secure when using BYOK frontier models?▼

Yes. All API keys (Google Gemini, Anthropic Claude, OpenAI) are saved strictly inside your browser's private localStorage. Requests are sent directly from your browser to the official provider API endpoints (generativelanguage.googleapis.com, api.anthropic.com, api.openai.com). No MCP Codex intermediary proxy ever touches your keys.

What MCP servers can I test in the browser studio?▼

You can test GitHub API (issues, PRs, repo search), PostgreSQL (SQL queries, table schemas), SQLite (local query execution), Brave Search (web queries), Filesystem sandboxes, Slack workspace integrations, Stripe payments, and custom remote SSE endpoints.

How does the MCP Protocol Inspector work?▼

The inspector captures every JSON-RPC 2.0 wire frame in real time. You can observe the exact 'tools/call' request payload, parameters, schema validation, raw response frames, and latency metrics (network roundtrip, tool execution time, and model inference latency).

Ready to Deploy MCP Servers in Production?

Use MCP Codex's visual generator to produce instant config snippets for Claude Desktop, Cursor, OpenAI Codex CLI, and Windsurf, or browse 30+ verified production servers.